Hello and welcome to Eye on AI. In this edition…new AI models are coming fast and furious; OpenAI CTO and cofounder Mira Murati departs as the company prepares to slip the shackles of its nonprofit parent; did Google just pay $2.7 billion largely to rehire one guy?
Over the last several days, Google, Meta, OpenAI, and AI2 all released new AI models. Weeks like this that are jam-packed with model releases used to be news-making events, but they’ve become almost hum drum as AI companies pour billions into developing and commercializing models as quickly as possible.
This pace of model releases says a lot about the current state of AI. It also makes it more difficult for customers—and everyone else—to keep up.
AI model releases become business as usual
If you feel like a new model is released practically every other day, you’re not wrong. When OpenAI released ChatGPT nearly two years ago, the company cracked the market for commercial AI models wide open. Now every player is working over time to release newer and better models, and they’ve started doing so at a headspinning pace.
This has its benefits. For one, customers have more options and that forces AI companies to compete on price and capabilities. On the other hand, having more options makes deciding which to tap more difficult, particularly because the technology is advancing so rapidly. In my conversations with chief technology officers and product managers building products based on LLMs, they’ve cited these types of technological decisions as a main challenge, saying they don’t want to lock themselves out of future progress with choices they make today.
As companies continue to pour resources into developing AI models, we can only expect this trend to continue. Some AI companies are working toward a specific goal (mainly AGI—which is artificial general intelligence, or AI as smart as a person—or “super intelligence,” an even more theoretical technology that would be smarter than all humans combined) and will keep releasing updates until they get there. Others are just trying to stay competitive. Perhaps the only exception is Ilya Sutskever’s Super Safe Intelligence (SSI), which vowed to not release anything until it succeeds in developing AGI.
What makes a model launch notable these days?
If you feel like a lot of the model releases these days just aren’t offering as much as they used to, you’re also not wrong about that. Many are new versions with incremental improvements. So what can we look for to decipher if a new model is notable? One factor would be if it offers use cases that are truly new.
Another signal to look out for is if a model offers similar capabilities but takes a very different approach. For example, currently the most capable models have come from companies going bigger and bigger—and some lawmakers are even taking model size, in terms of the number of parameters it has, as the metric to determine which regulations it’s subject to. Of course, bigger models have huge drawbacks around the amount of compute and resources they require.
But another trend has been companies trying to match some of the capabilities of the larger models but with much smaller ones. That's the case with AI2's newest model family called Molmo, which it says can match or exceed OpenAI's GPT-4o even though the models are a fraction of the size. To the extent companies are able to achieve these milestones, that’s quite significant.
And with that, here’s more AI news.
Sage Lazzaro
sage.lazzaro@consultant.fortune.com
sagelazzaro.com
